Parallel Programming

1
Parallel Programming Cluster ComputingDistribut
ed Parallelism

• Henry Neeman, University of Oklahoma
• Paul Gray, University of Northern Iowa
• SC08 Education Programs Workshop on Parallel
  Cluster Computing
• Oklahoma Supercomputing Symposium, Monday October
  6 2008

2
Outline

• The Desert Islands Analogy
• Distributed Parallelism
• MPI

3
The Desert Islands Analogy
4
An Island Hut

• Imagine youre on an island in a little hut.
• Inside the hut is a desk.
• On the desk is
• a phone
• a pencil
• a calculator
• a piece of paper with numbers
• a piece of paper with instructions.

5
Instructions

• The instructions are split into two kinds
• Arithmetic/Logical e.g.,
• Add the 27th number to the 239th number
• Compare the 96th number to the 118th number to
  see whether they are equal
• Communication e.g.,
• dial 555-0127 and leave a voicemail containing
  the 962nd number
• call your voicemail box and collect a voicemail
  from 555-0063 and put that number in the 715th
  slot

6
Is There Anybody Out There?

• If youre in a hut on an island, you arent
  specifically aware of anyone else.
• Especially, you dont know whether anyone else is
  working on the same problem as you are, and you
  dont know whos at the other end of the phone
  line.
• All you know is what to do with the voicemails
  you get, and what phone numbers to send
  voicemails to.

7
Someone Might Be Out There

• Now suppose that Horst is on another island
  somewhere, in the same kind of hut, with the same
  kind of equipment.
• Suppose that he has the same list of instructions
  as you, but a different set of numbers (both data
  and phone numbers).
• Like you, he doesnt know whether theres anyone
  else working on his problem.

8
Even More People Out There

• Now suppose that Bruce and Dee are also in huts
  on islands.
• Suppose that each of the four has the exact same
  list of instructions, but different lists of
  numbers.
• And suppose that the phone numbers that people
  call are each others. That is, your
  instructions have you call Horst, Bruce and Dee,
  Horsts has him call Bruce, Dee and you, and so
  on.
• Then you might all be working together on the
  same problem.

9
All Data Are Private

• Notice that you cant see Horsts or Bruces or
  Dees numbers, nor can they see yours or each
  others.
• Thus, everyones numbers are private theres no
  way for anyone to share numbers, except by
  leaving them in voicemails.

10
Long Distance Calls 2 Costs

• When you make a long distance phone call, you
  typically have to pay two costs
• Connection charge the fixed cost of connecting
  your phone to someone elses, even if youre only
  connected for a second
• Per-minute charge the cost per minute of
  talking, once youre connected
• If the connection charge is large, then you want
  to make as few calls as possible.

11
Like Desert Islands

• Distributed parallelism is very much like the
  Desert Islands analogy
• Processors are independent of each other.
• All data are private.
• Processes communicate by passing messages (like
  voicemails).
• The cost of passing a message is split into the
  latency (connection time) and the bandwidth (time
  per byte).

12
Latency vs Bandwidth on topdawg

• In 2006, we benchmarked the Infiniband
  interconnect on OUs large Linux cluster
  (topdawg.oscer.ou.edu).
• Latency the time for the first bit to show up
  at the destination is about 3 microseconds
• Bandwidth the speed of the subsequent bits is
  about 5 Gigabits per second.
• Thus, on topdawgs Infiniband
• the 1st bit of a message shows up in 3 microsec
• the 2nd bit shows up in 0.2 nanosec.
• So latency is 15,000 times worse than bandwidth!

13
Latency vs Bandwidth on topdawg

• In 2006, we benchmarked the Infiniband
  interconnect on OUs large Linux cluster
  (topdawg.oscer.ou.edu).
• Latency the time for the first bit to show up
  at the destination is about 3 microseconds
• Bandwidth the speed of the subsequent bits is
  about 5 Gigabits per second.
• Latency is 15,000 times worse than bandwidth!
• Thats like having a long distance service that
  charges
• 150 to make a call
• 1 per minute after the first 10 days of the
  call.

14
Distributed Parallelism
15
Like Desert Islands

• Distributed parallelism is very much like the
  Desert Islands analogy
• processes are independent of each other.
• All data are private.
• Processes communicate by passing messages (like
  voicemails).
• The cost of passing a message is split into
• latency (connection time)
• bandwidth (time per byte)

16
Parallelism
Parallelism means doing multiple things at the
same time you can get more work done in the same
amount of time.
Less fish
More fish!
17
What Is Parallelism?

• Parallelism is the use of multiple processing
  units either processors or parts of an
  individual processor to solve a problem, and in
  particular the use of multiple processing units
  operating concurrently on different parts of a
  problem.
• The different parts could be different tasks, or
  the same task on different pieces of the
  problems data.

18
Kinds of Parallelism

• Shared Memory Multithreading (our topic last
  time)
• Distributed Memory Multiprocessing (today)
• Hybrid Shared/Distributed

19
Why Parallelism Is Good

• The Trees We like parallelism because, as the
  number of processing units working on a problem
  grows, we can solve the same problem in less
  time.
• The Forest We like parallelism because, as the
  number of processing units working on a problem
  grows, we can solve bigger problems.

20
Parallelism Jargon

• Threads execution sequences that share a single
  memory area (address space)
• Processes execution sequences with their own
  independent, private memory areas
• and thus
• Multithreading parallelism via multiple
  threads
• Multiprocessing parallelism via multiple
  processes
• As a general rule, Shared Memory Parallelism is
  concerned with threads, and Distributed
  Parallelism is concerned with processes.

21
Jargon Alert

• In principle
• shared memory parallelism ? multithreading
• distributed parallelism ?
  multiprocessing
• In practice, these terms are often used
  interchangeably
• Parallelism
• Concurrency (not as popular these days)
• Multithreading
• Multiprocessing
• Typically, you have to figure out what is meant
  based on the context.

22
Load Balancing

• Suppose you have a distributed parallel code, but
  one process does 90 of the work, and all the
  other processes share 10 of the work.
• Is it a big win to run on 1000 processes?
• Now, suppose that each process gets exactly 1/Np
  of the work, where Np is the number of processes.
• Now is it a big win to run on 1000 processes?

23
Load Balancing
Load balancing means giving everyone roughly the
same amount of work to do.
24
Load Balancing
Load balancing can be easy, if the problem splits
up into chunks of roughly equal size, with one
chunk per process. Or load balancing can be very
hard.
25
Load Balancing Is Good

• When every process gets the same amount of work,
  the job is load balanced.
• We like load balancing, because it means that our
  speedup can potentially be linear if we run on
  Np processes, it takes 1/Np as much time as on
  one.
• For some codes, figuring out how to balance the
  load is trivial (e.g., breaking a big unchanging
  array into sub-arrays).
• For others, load balancing is very tricky (e.g.,
  a dynamically evolving collection of arbitrarily
  many blocks of arbitrary size).

26
Parallel Strategies

• Client-Server One worker (the server) decides
  what tasks the other workers (clients) will do
  e.g., Hello World, Monte Carlo.
• Data Parallelism Each worker does exactly the
  same tasks on its unique subset of the data
  e.g., distributed meshes (weather etc).
• Task Parallelism Each worker does different
  tasks on exactly the same set of data (each
  process holds exactly the same data as the
  others) e.g., N-body.
• Pipeline Each worker does its tasks, then passes
  its set of data along to the next worker and
  receives the next set of data from the previous
  worker.

27
MPIThe Message-Passing Interface
Most of this discussion is from 1 and 2.
28
What Is MPI?

• The Message-Passing Interface (MPI) is a standard
  for expressing distributed parallelism via
  message passing.
• MPI consists of a header file, a library of
  routines and a runtime environment.
• When you compile a program that has MPI calls in
  it, your compiler links to a local implementation
  of MPI, and then you get parallelism if the MPI
  library isnt available, then the compile will
  fail.
• MPI can be used in Fortran, C and C.

29
MPI Calls

• MPI calls in Fortran look like this
• CALL MPI_Funcname(, errcode)
• In C, MPI calls look like
• errcode MPI_Funcname()
• In C, MPI calls look like
• errcode MPIFuncname()
• Notice that errcode is returned by the MPI
  routine MPI_Funcname, with a value of MPI_SUCCESS
  indicating that MPI_Funcname has worked correctly.

30
MPI is an API

• MPI is actually just an Application Programming
  Interface (API).
• An API specifies what a call to each routine
  should look like, and how each routine should
  behave.
• An API does not specify how each routine should
  be implemented, and sometimes is intentionally
  vague about certain aspects of a routines
  behavior.
• Each platform has its own MPI implementation.

31
Example MPI Routines

• MPI_Init starts up the MPI runtime environment at
  the beginning of a run.
• MPI_Finalize shuts down the MPI runtime
  environment at the end of a run.
• MPI_Comm_size gets the number of processes in a
  run, Np (typically called just after MPI_Init).
• MPI_Comm_rank gets the process ID that the
  current process uses, which is between 0 and Np-1
  inclusive (typically called just after MPI_Init).

32
More Example MPI Routines

• MPI_Send sends a message from the current process
  to some other process (the destination).
• MPI_Recv receives a message on the current
  process from some other process (the source).
• MPI_Bcast broadcasts a message from one process
  to all of the others.
• MPI_Reduce performs a reduction (e.g., sum,
  maximum) of a variable on all processes, sending
  the result to a single process.

33
MPI Program Structure (F90)

• PROGRAM my_mpi_program
• IMPLICIT NONE
• INCLUDE "mpif.h"
• other includes
• INTEGER my_rank, num_procs, mpi_error_code
• other declarations
• CALL MPI_Init(mpi_error_code) !! Start up
  MPI
• CALL MPI_Comm_Rank(my_rank, mpi_error_code)
• CALL MPI_Comm_size(num_procs, mpi_error_code)
• actual work goes here
• CALL MPI_Finalize(mpi_error_code) !! Shut down
  MPI
• END PROGRAM my_mpi_program
• Note that MPI uses the term rank to indicate
  process identifier.

34
MPI Program Structure (in C)

• include ltstdio.hgt
• include "mpi.h"
• other includes
• int main (int argc, char argv)
• / main /
• int my_rank, num_procs, mpi_error
• other declarations
• mpi_error MPI_Init(argc, argv) / Start up
  MPI /
• mpi_error MPI_Comm_rank(MPI_COMM_WORLD,
  my_rank)
• mpi_error MPI_Comm_size(MPI_COMM_WORLD,
  num_procs)
• actual work goes here
• mpi_error MPI_Finalize() / Shut
  down MPI /
• / main /

35
MPI is SPMD

• MPI uses kind of parallelism known as Single
  Program, Multiple Data (SPMD).
• This means that you have one MPI program a
  single executable that is executed by all of
  the processes in an MPI run.
• So, to differentiate the roles of various
  processes in the MPI run, you have to have if
  statements
• if (my_rank server_rank)

36
Example Hello World

• Start the MPI system.
• Get the rank and number of processes.
• If youre not the server process
• Create a hello world string.
• Send it to the server process.
• If you are the server process
• For each of the client processes
• Receive its hello world string.
• Print its hello world string.
• Shut down the MPI system.

37
hello_world_mpi.c

• include ltstdio.hgt
• include ltstring.hgt
• include "mpi.h"
• int main (int argc, char argv)
• / main /
• const int maximum_message_length 100
• const int server_rank 0
• char messagemaximum_message_length1
• MPI_Status status / Info about receive
  status /
• int my_rank / This process ID
  /
• int num_procs / Number of processes
  in run /
• int source / Process ID to
  receive from /
• int destination / Process ID to send
  to /
• int tag 0 / Message ID
  /
• int mpi_error / Error code for MPI
  calls /
• work goes here
• / main /

38
Hello World Startup/Shut Down

• header file includes
• int main (int argc, char argv)
• / main /
• declarations
• mpi_error MPI_Init(argc, argv)
• mpi_error MPI_Comm_rank(MPI_COMM_WORLD,
  my_rank)
• mpi_error MPI_Comm_size(MPI_COMM_WORLD,
  num_procs)
• if (my_rank ! server_rank)
• work of each non-server (worker)
  process
• / if (my_rank ! server_rank) /
• else
• work of server process
• / if (my_rank ! server_rank)else /
• mpi_error MPI_Finalize()
• / main /

39
Hello World Clients Work

• header file includes
• int main (int argc, char argv)
• / main /
• declarations
• MPI startup (MPI_Init etc)
• if (my_rank ! server_rank)
• sprintf(message, "Greetings from process
  d!,
• my_rank)
• destination server_rank
• mpi_error
• MPI_Send(message, strlen(message) 1,
  MPI_CHAR,
• destination, tag, MPI_COMM_WORLD)
• / if (my_rank ! server_rank) /
• else
• work of server process
• / if (my_rank ! server_rank)else /
• mpi_error MPI_Finalize()
• / main /

40
Hello World Servers Work

• header file includes
• int main (int argc, char argv)
• / main /
• declarations, MPI startup
• if (my_rank ! server_rank)
• work of each client process
• / if (my_rank ! server_rank) /
• else
• for (source 0 source lt num_procs
  source)
• if (source ! server_rank)
• mpi_error
• MPI_Recv(message, maximum_message_length
  1,
• MPI_CHAR, source, tag,
  MPI_COMM_WORLD,
• status)
• fprintf(stderr, "s\n", message)
• / if (source ! server_rank) /
• / for source /
• / if (my_rank ! server_rank)else /
• mpi_error MPI_Finalize()

41
How an MPI Run Works

• Every process gets a copy of the executable
  Single Program, Multiple Data (SPMD).
• They all start executing it.
• Each looks at its own rank to determine which
  part of the problem to work on.
• Each process works completely independently of
  the other processes, except when communicating.

42
Compiling and Running

• mpicc -o hello_world_mpi hello_world_mpi.c
• mpirun -np 1 hello_world_mpi
• mpirun -np 2 hello_world_mpi
• Greetings from process 1!
• mpirun -np 3 hello_world_mpi
• Greetings from process 1!
• Greetings from process 2!
• mpirun -np 4 hello_world_mpi
• Greetings from process 1!
• Greetings from process 2!
• Greetings from process 3!
• Note The compile command and the run command
  vary from platform to platform.

43
Why is Rank 0 the server?

• const int server_rank 0
• By convention, the server process has rank
  (process ID) 0. Why?
• A run must use at least one process but can use
  multiple processes.
• Process ranks are 0 through Np-1, Np gt1 .
• Therefore, every MPI run has a process with rank
  0.
• Note Every MPI run also has a process with rank
  Np-1, so you could use Np-1 as the server instead
  of 0 but no one does.

44
Why Rank?

• Why does MPI use the term rank to refer to
  process ID?
• In general, a process has an identifier that is
  assigned by the operating system (e.g., Unix),
  and that is unrelated to MPI
• ps
• PID TTY TIME CMD
• 52170812 ttyq57 001 tcsh
• Also, each processor has an identifier, but an
  MPI run that uses fewer than all processors will
  use an arbitrary subset.
• The rank of an MPI process is neither of these.

45
Compiling and Running

• Recall
• mpicc -o hello_world_mpi hello_world_mpi.c
• mpirun -np 1 hello_world_mpi
• mpirun -np 2 hello_world_mpi
• Greetings from process 1!
• mpirun -np 3 hello_world_mpi
• Greetings from process 1!
• Greetings from process 2!
• mpirun -np 4 hello_world_mpi
• Greetings from process 1!
• Greetings from process 2!
• Greetings from process 3!

46
Deterministic Operation?

• mpirun -np 4 hello_world_mpi
• Greetings from process 1!
• Greetings from process 2!
• Greetings from process 3!
• The order in which the greetings are printed is
  deterministic. Why?
• for (source 0 source lt num_procs source)
• if (source ! server_rank)
• mpi_error
• MPI_Recv(message, maximum_message_length
  1,
• MPI_CHAR, source, tag, MPI_COMM_WORLD,
• status)
• fprintf(stderr, "s\n", message)
• / if (source ! server_rank) /
• / for source /
• This loop ignores the receive order.

47
Message EnvelopeContents

• MPI_Send(message, strlen(message) 1,
• MPI_CHAR, destination, tag,
• MPI_COMM_WORLD)
• When MPI sends a message, it doesnt just send
  the contents it also sends an envelope
  describing the contents
• Size (number of elements of data type)
• Data type
• Source rank of sending process
• Destination rank of process to receive
• Tag (message ID)
• Communicator (e.g., MPI_COMM_WORLD)

48
MPI Data Types
MPI supports several other data types, but most
are variations of these, and probably these are
all youll use.
49
Message Tags

• for (source 0 source lt num_procs source)
• if (source ! server_rank)
• mpi_error
• MPI_Recv(message, maximum_message_length
  1,
• MPI_CHAR, source, tag,
• MPI_COMM_WORLD, status)
• fprintf(stderr, "s\n", message)
• / if (source ! server_rank) /
• / for source /
• The greetings are printed in deterministic
  order not because messages are sent and received
  in order, but because each has a tag (message
  identifier), and MPI_Recv asks for a specific
  message (by tag) from a specific source (by rank).

50
Parallelism is Nondeterministic

• for (source 0 source lt num_procs source)
• if (source ! server_rank)
• mpi_error
• MPI_Recv(message, maximum_message_length
  1,
• MPI_CHAR, MPI_ANY_SOURCE, tag,
• MPI_COMM_WORLD, status)
• fprintf(stderr, "s\n", message)
• / if (source ! server_rank) /
• / for source /
• The greetings are printed in non-deterministic
  order.

51
Communicators

• An MPI communicator is a collection of processes
  that can send messages to each other.
• MPI_COMM_WORLD is the default communicator it
  contains all of the processes. Its probably the
  only one youll need.
• Some libraries create special library-only
  communicators, which can simplify keeping track
  of message tags.

52
Broadcasting

• What happens if one process has data that
  everyone else needs to know?
• For example, what if the server process needs to
  send an input value to the others?
• MPI_Bcast(length, 1, MPI_INTEGER,
• source, MPI_COMM_WORLD)
• Note that MPI_Bcast doesnt use a tag, and that
  the call is the same for both the sender and all
  of the receivers.
• All processes have to call MPI_Bcast at the same
  time everyone waits until everyone is done.

53
Broadcast Example Setup

• PROGRAM broadcast
• IMPLICIT NONE
• INCLUDE "mpif.h"
• INTEGER,PARAMETER server 0
• INTEGER,PARAMETER source server
• INTEGER,DIMENSION(),ALLOCATABLE array
• INTEGER length, memory_status
• INTEGER num_procs, my_rank, mpi_error_code
• CALL MPI_Init(mpi_error_code)
• CALL MPI_Comm_rank(MPI_COMM_WORLD, my_rank,
• mpi_error_code)
• CALL MPI_Comm_size(MPI_COMM_WORLD, num_procs,
• mpi_error_code)
• input
• broadcast
• CALL MPI_Finalize(mpi_error_code)
• END PROGRAM broadcast

54
Broadcast Example Input

• PROGRAM broadcast
• IMPLICIT NONE
• INCLUDE "mpif.h"
• INTEGER,PARAMETER server 0
• INTEGER,PARAMETER source server
• INTEGER,DIMENSION(),ALLOCATABLE array
• INTEGER length, memory_status
• INTEGER num_procs, my_rank, mpi_error_code
• MPI startup
• IF (my_rank server) THEN
• OPEN (UNIT99,FILE"broadcast_in.txt")
• READ (99,) length
• CLOSE (UNIT99)
• ALLOCATE(array(length), STATmemory_status)
• array(1length) 0
• END IF !! (my_rank server)...ELSE
• broadcast
• CALL MPI_Finalize(mpi_error_code)

55
Broadcast Example Broadcast

• PROGRAM broadcast
• IMPLICIT NONE
• INCLUDE "mpif.h"
• INTEGER,PARAMETER server 0
• INTEGER,PARAMETER source server
• other declarations
• MPI startup and input
• IF (num_procs gt 1) THEN
• CALL MPI_Bcast(length, 1, MPI_INTEGER,
  source,
• MPI_COMM_WORLD, mpi_error_code)
• IF (my_rank / server) THEN
• ALLOCATE(array(length), STATmemory_status)
• END IF !! (my_rank / server)
• CALL MPI_Bcast(array, length, MPI_INTEGER,
  source,
• MPI_COMM_WORLD, mpi_error_code)
• WRITE (0,) my_rank, " broadcast length ",
  length
• END IF !! (num_procs gt 1)
• CALL MPI_Finalize(mpi_error_code)

56
Broadcast Compile Run

• mpif90 -o broadcast broadcast.f90
• mpirun -np 4 broadcast
• 0 broadcast length 16777216
• 1 broadcast length 16777216
• 2 broadcast length 16777216
• 3 broadcast length 16777216

57
Reductions

• A reduction converts an array to a scalar for
  example, sum, product, minimum value,
  maximum value, Boolean AND, Boolean OR, etc.
• Reductions are so common, and so important, that
  MPI has two routines to handle them
• MPI_Reduce sends result to a single specified
  process
• MPI_Allreduce sends result to all processes (and
  therefore takes longer)

58
Reduction Example

• PROGRAM reduce
• IMPLICIT NONE
• INCLUDE "mpif.h"
• INTEGER,PARAMETER server 0
• INTEGER value, value_sum
• INTEGER num_procs, my_rank, mpi_error_code
• CALL MPI_Init(mpi_error_code)
• CALL MPI_Comm_rank(MPI_COMM_WORLD, my_rank,
  mpi_error_code)
• CALL MPI_Comm_size(MPI_COMM_WORLD, num_procs,
  mpi_error_code)
• value_sum 0
• value my_rank num_procs
• CALL MPI_Reduce(value, value_sum, 1, MPI_INT,
  MPI_SUM,
• server, MPI_COMM_WORLD, mpi_error_code)
• WRITE (0,) my_rank, " reduce value_sum ",
  value_sum
• CALL MPI_Allreduce(value, value_sum, 1,
  MPI_INT, MPI_SUM,
• MPI_COMM_WORLD, mpi_error_code)
• WRITE (0,) my_rank, " allreduce value_sum
  ", value_sum
• CALL MPI_Finalize(mpi_error_code)

59
Compiling and Running

• mpif90 -o reduce reduce.f90
• mpirun -np 4 reduce
• 3 reduce value_sum 0
• 1 reduce value_sum 0
• 2 reduce value_sum 0
• 0 reduce value_sum 24
• 0 allreduce value_sum 24
• 1 allreduce value_sum 24
• 2 allreduce value_sum 24
• 3 allreduce value_sum 24

60
Why Two Reduction Routines?

• MPI has two reduction routines because of the
  high cost of each communication.
• If only one process needs the result, then it
  doesnt make sense to pay the cost of sending the
  result to all processes.
• But if all processes need the result, then it may
  be cheaper to reduce to all processes than to
  reduce to a single process and then broadcast to
  all.

61
Non-blocking Communication

• MPI allows a process to start a send, then go on
  and do work while the message is in transit.
• This is called non-blocking or immediate
  communication.
• Here, immediate refers to the fact that the
  call to the MPI routine returns immediately
  rather than waiting for the communication to
  complete.

62
Immediate Send

• mpi_error_code
• MPI_Isend(array, size, MPI_FLOAT,
• destination, tag, communicator, request)
• Likewise
• mpi_error_code
• MPI_Irecv(array, size, MPI_FLOAT,
• source, tag, communicator, request)
• This call starts the send/receive, but the
  send/receive wont be complete until
• MPI_Wait(request, status)
• Whats the advantage of this?

63
Communication Hiding

• In between the call to MPI_Isend/Irecv and the
  call to MPI_Wait, both processes can do work!
• If that work takes at least as much time as the
  communication, then the cost of the communication
  is effectively zero, since the communication
  wont affect how much work gets done.
• This is called communication hiding.

64
Rule of Thumb for Hiding

• When you want to hide communication
• as soon as you calculate the data, send it
• dont receive it until you need it.
• That way, the communication has the maximal
  amount of time to happen in background (behind
  the scenes).

65
To Learn More Supercomputing

• http//www.oscer.ou.edu/education.php

66
Thanks for your attention!Questions?
67
References
1 P.S. Pacheco, Parallel Programming with MPI,
Morgan Kaufmann Publishers, 1997. 2 W.
Gropp, E. Lusk and A. Skjellum, Using MPI
Portable Parallel Programming with the
Message-Passing Interface, 2nd ed. MIT
Press, 1999.